基于改进SSD的棉种破损检测

Improved SSD based detection of damaged cottonseed

为实现群体棉籽的破损检测,以新路早-50#脱绒棉籽为研究对象,将群体棉籽随机摆放,使用CCD相机采集群体棉籽的图像,在经典的单步多框检测(single shot multibox detector,SSD)算法上进行改进。基于改进SSD,利用ResNet 50网络代替经典SSD算法中的VGG网络,将ResNet50作为SSD的基础网络,用来快速提取群体棉籽图像的特征,最终对群体脱绒棉籽中的破损棉籽实现精准识别。试验结果表明:该方法建立的模型对群体棉籽的检测精度、召回率、漏检率分别达到96.1%、97.3%、0%;高于经典SSD网络模型(检测精度、召回率、漏检率分别为92.5%、96.4%、1.4%)。

Cotton production in China is huge.The southern region of Xinjiang is one important cotton production base of China.At the same time,cotton production is a pillar industry in this region as well.However,the quality and sorting problem of lint-free cotton has severely limited the development of the cotton industry in this region.In order to realize the detection of damaged population cottonseed,the population cottonseed of Xinluzao-50#lint-free cottonseed were randomly arranged.The CCD camera was used to collect the image of the population cottonseed.The classic single-step multi-frame detection(single shot multibox detector,SSD)algorithm was improved.Based on the improved SSD,the ResNet50 network was used to replace the VGG network in the classic SSD algorithm.ResNet50 was used as the basic network of the SSD to quickly extract the image characteristics of the population cottonseed,and to finally realize the accurate identification of the damaged cottonseed in the population lint-free cottonseed.The results showed that the detection accuracy,recall rate,and missed detection rate of the model established by this method for the damaged cottonseed and the non-destructive cottonseed in the population cottonseed was 96.1%,97.3%,and 0%,respectively.It is higher than that(92.5%,96.4%,1.4%)of the classic SSD network model.This study transfers the pre-trained model weights under the COCO large dataset to the task of damage and non-destructive detection in the population cottonseed,which accelerates the convergence speed of the population cottonseed detection model and saves the training cost of the model.It solves the problem of the difficulty of segmentation of the population cottonseed image.It directly uses the convolutional neural network to obtain the position and category information of the cottonseed.It is not necessary to use traditional image recognition methods to separate the individuals in population cottonseed for detection.It will provide a novel idea for detecting population cottonseed damage to accelerate the intelligent sorting of cottonseed and a technical support for subsequently studying and developing related automation equipment.